Chinese giant releases open weights for 2.4 trillion parameter MoE model

Alibaba has released the open weights for Qwen3.8-2.4T-A95B, a Mixture of Experts model with 2.4 trillion total parameters and 95 billion active parameters per forward pass. The announcement came via Nvidia's AI infrastructure account, which noted the model was designed for demanding reasoning and agentic workloads.

Architecture and numbers

In an MoE architecture, only a subset of parameters activates per token, allowing total capacity to scale without a proportional increase in compute cost. Here the ratio stands at roughly 25:1 between total and active parameters, placing the model among the largest open-weight models published to date. Nvidia did not specify how many experts the model contains or its context window size, and the accompanying technical blog has not yet been fully analyzed publicly.

Performance on GB300 NVL72 hardware

According to manufacturer data, on an Nvidia GB300 NVL72 system at FP8 precision the model achieves 4,000 tokens per second per GPU and 350 tokens per second per user. The figures refer to inference only, not training, and were measured "out of the box" — without additional optimizations such as advanced quantization or speculative decoding. The source provides no comparison benchmarks against peer models such as Nvidia's own Nemotron 3 Ultra or earlier Qwen versions.

Community response: workstation-scale compression

In parallel with the announcement, an independent researcher reported compressing Kimi K3, another large Chinese model, to run on three Nvidia DGX Spark workstations at 12.5 tokens per second. Published benchmarks: 94.5% on HumanEval, 92.5% on AIME, 95% on GSM8K, and 79.5% on MMLU. Results approach those of the full model, but the sample is small and the tests have not undergone external validation.

What's missing from the picture

No performance metrics on complex reasoning tasks such as GPQA or LiveCodeBench have been published. There is no discussion of safety or alignment, and it remains unclear whether the weights carry a fully commercial license or an "open weights" model with usage restrictions. The final training date (knowledge cutoff) was also not specified. Until the technical blog is analyzed and the community runs independent evaluations, this remains an impressive-on-paper announcement, not a proven capability in production.